Curvature Diversity-Driven Deformation and Domain Alignment for Point Cloud

TL;DR

The proposed CDND achieves state-of-the-art performance by a noticeable margin over existing approaches and contributes a theoretical justification for the effectiveness of D-NWD in distribution alignment.

Abstract

Unsupervised Domain Adaptation (UDA) is crucial for reducing the need for extensive manual data annotation when training deep networks on point cloud data. A significant challenge of UDA lies in effectively bridging the domain gap. To tackle this challenge, we propose \textbf{C}urvature \textbf{D}iversity-Driven \textbf{N}uclear-Norm Wasserstein \textbf{D}omain Alignment (CDND). Our approach first introduces a \textit{\textbf{Curv}ature Diversity-driven Deformation \textbf{Rec}onstruction (CurvRec)} task, which effectively mitigates the gap between the source and target domains by enabling the model to extract salient features from semantically rich regions of a given point cloud. We then propose \textit{\textbf{D}eformation-based \textbf{N}uclear-norm \textbf{W}asserstein \textbf{D}iscrepancy (D-NWD)}, which applies the Nuclear-norm Wasserstein Discrepancy to both \textit{deformed and original} data samples to align the source and target domains. Furthermore, we contribute a theoretical justification for the effectiveness of D-NWD in distribution alignment and demonstrate that it is \textit{generic} enough to be applied to \textbf{any} deformations. To validate our method, we conduct extensive experiments on two public domain adaptation datasets for point cloud classification and segmentation tasks. Empirical experiment results show that our CDND achieves state-of-the-art performance by a noticeable margin over existing approaches.

Figures (4)

• Figure 1: Pipeline of CDND. The inputs are the source batch $X_s$ and target batch $X_t$. We first deform them into $\hat{X}_s$ and $\hat{X}_t$ using Curvature Diversity-Based Deformation. Next, $X_s$, $X_t$, $\hat{X}_s$, and $\hat{X}_t$ are sent into a feature extractor. The features of deformed samples are fed into a reconstruction decoder to reconstruct the deformed regions. For domain alignment, both original and deformed features are sent to D-NWD. Aside from the two losses shown in the figure, a cross-entropy loss is computed on $X_s$ and $\hat{X}_s$ with labels. An NWD loss $\mathcal{L}_{\text{NWD}}^{\mathcal{T}}$ on $X_t$ and $\hat{X}_t$ is also computed to ensure prediction consistency between the target original and deformed pairs.
• Figure 2: UMAP visualizations depict pre-activation data representations for the MS+ task, with different colors denoting different classes. The center plot shows the target domain test data representations generated from a model trained on the source dataset without any adaptation. The left and right plots show the source and target domain data representations after adaptation using CDND.
• Figure 3: Trianing curve for ScanNet to ShapeNet task from PointDA.
• Figure 4: The performance of CDND with different hyperparameter settings on PointSegDA dataset.